.gitattributes

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+Qwen3-4B-Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text

1/model.py

@@ -1,260 +0,0 @@
-"""
-[Transformer-LLM 백엔드 가이드]
-
-본 파일은 NVIDIA Triton Server에서 Hugging Face `AutoModelForCausalLM` 기반 모델을 손쉽게 배포하기 위해 제공되는 커스텀 Python 백엔드 템플릿입니다.
-
-1. 모델 호환성
-   - Hugging Face의 `AutoModelForCausalLM` 클래스와 호환되는 모든 Causal Language Model을 지원합니다.
-   - [확인] 배포할 모델 `config.json`의 `architectures` 항목이 `...ForCausalLM` 형식인지 확인.
-   - [확인] 모델이 Hugging Face 공식 문서의 AutoModelForCausalLM이 지원하는 모델인지 확인.
-           (https://huggingface.co/docs/transformers/en/model_doc/auto#transformers.AutoModelForCausalLM.from_pretrained)
-
-2. 토크나이저 호환성
-   - `AutoTokenizer`와 호환되는 토크나이저를 지원하며, 모델과 동일한 경로에서 자동으로 로드됩니다.
-
-3. 커스터마이징 안내
-   - 본 템플릿은 범용적인 사용을 위해 작성되었습니다.
-   - 특정 모델의 동작 방식이나 예외 처리가 필요한 경우, 이 파일(`model.py`)과 설정 파일(`config.pbtxt`)을 직접 수정하여 사용하시기 바랍니다.
-"""
-
-import json
-import torch
-import numpy as np
-import triton_python_backend_utils as pb_utils
-import uuid
-import transformers
-from typing import List, Dict, Any, Union, Tuple
-
-from transformers import (
-    AutoModelForCausalLM,
-    AutoTokenizer,
-    GenerationConfig,
-    BitsAndBytesConfig,
-)
-from peft import PeftModel, PeftConfig
-
-class TritonPythonModel:
-    def initialize(self, args: Dict[str, str]):
-        """
-        모델 초기화: 라이브러리 버전 확인 및 모델/토크나이저 로드
-        """
-        self.logger = pb_utils.Logger
-        self.model_config = json.loads(args["model_config"])
-        self.model_name = args["model_name"]
-
-        # 1. 라이브러리 버전 로그 추가
-        # GGUF 로드를 위해서는 최소 4.40.0 이상을 권장합니다.
-        transformers_version = transformers.__version__
-        self.logger.log_info(f"================ {self.model_name} Setup ================")
-        self.logger.log_info(f"Transformers Version: {transformers_version}")
-        self.logger.log_info(f"Torch Version: {torch.__version__}")
-
-        # 설정 파라미터 로드
-        self.base_model_path = self._get_config_param("base_model_path")
-        self.gguf_filename = self._get_config_param("gguf_filename")
-        self.device = "cuda" if torch.cuda.is_available() else "cpu"
-
-        self.logger.log_info(f"Base Model Path: {self.base_model_path}")
-        self.logger.log_info(f"GGUF Filename: {self.gguf_filename}")
-        self.logger.log_info(f"Device: {self.device}")
-
-        # 2. 모델 및 토크나이저 로드 실행
-        self._load_model_and_tokenizer()
-        self.logger.log_info(f"Model initialized successfully.")
-
-    def _load_model_and_tokenizer(self):
-        """
-        config.pbtxt의 파라미터를 사용하여 GGUF 모델을 로드합니다.
-        Transformers 라이브러리가 GGUF를 읽어 fp16으로 역양자화합니다.
-        """
-        # 1. config.pbtxt에서 설정값 읽기
-        load_path = self.base_model_path  # /cheetah/input/model/groupuser/Qwen3-4B-Instruct-2507-mahjong-alpha
-        gguf_file = self._get_config_param("gguf_filename") # Qwen3-4B-Instruct-2507-mahjong-alpha.gguf
-
-        self.logger.log_info(f"Loading GGUF from: {load_path}/{gguf_file}")
-
-        try:
-            # 2. Tokenizer 로드 (GGUF 파일 내의 토크나이저 메타데이터 참조)
-            self.tokenizer = AutoTokenizer.from_pretrained(
-                load_path, 
-                gguf_file=gguf_file,
-                trust_remote_code=True
-            )
-
-            # 3. Model 로드 (GGUF -> PyTorch fp16 변환)
-            # 주의: GGUF 로드 시 bnb_config(int4/8)와 중복 사용은 불가능할 수 있습니다.
-            self.model = AutoModelForCausalLM.from_pretrained(
-                load_path,
-                gguf_file=gguf_file,
-                torch_dtype=torch.float16,
-                device_map="auto",
-                local_files_only=True,
-                trust_remote_code=True
-            )
-            
-            self.model.eval()
-
-            # 패딩 토큰 설정
-            if self.tokenizer.pad_token is None:
-                self.tokenizer.pad_token = self.tokenizer.eos_token
-                self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
-
-            self.supports_chat_template = (
-                hasattr(self.tokenizer, "chat_template") and
-                self.tokenizer.chat_template is not None
-            )
-
-            self.logger.log_info("GGUF Model and Tokenizer loaded successfully via Transformers.")
-
-        except Exception as e:
-            self.logger.log_error(f"Failed to load GGUF model: {e}")
-            raise e
-
-    def _get_bnb_config(self) -> Union[BitsAndBytesConfig, None]:
-        if self.quantization == "int4":
-            return BitsAndBytesConfig(
-                load_in_4bit=True,
-                bnb_4bit_use_double_quant=True,
-                bnb_4bit_quant_type="nf4",
-                bnb_4bit_compute_dtype=torch.float16
-            )
-        elif self.quantization == "int8":
-            return BitsAndBytesConfig(
-                load_in_8bit=True,
-                llm_int8_threshold=6.0,
-                llm_int8_has_fp16_weight=True
-            )
-        return None
-
-    def execute(self, requests):
-        """Triton Inference Request 처리 메인 루프"""
-        responses = []
-
-        for request in requests:
-            # [ID 생성 로직] - 로그 추적용으로 유지 (Response에는 포함 X)
-            request_id = request.request_id()
-            if not request_id:
-                request_id = str(uuid.uuid4())
-
-            try:
-                # 1. 입력 데이터 파싱
-                input_data, is_chat = self._parse_input(request)
-
-                # [LOGGING] Request ID 포함하여 로그 출력
-                log_input_str = json.dumps(input_data, ensure_ascii=False) if isinstance(input_data, (list, dict)) else str(input_data)
-                self.logger.log_info(f"\n[RID: {request_id}] >>> [{'CHAT' if is_chat else 'TEXT'}][Input]: {log_input_str}")
-
-                # 2. Generation Config 생성
-                gen_config = self._create_generation_config(request)
-
-                # 3. 토크나이징
-                inputs = self._tokenize(input_data, is_chat)
-
-                # 4. 모델 추론 (Generate)
-                output_text = self._generate(inputs, gen_config)
-
-                # [LOGGING] Request ID 포함하여 결과 출력
-                self.logger.log_info(f"\n[RID: {request_id}] <<< [Output]: {output_text}")
-
-                # 5. 응답 생성
-                responses.append(self._create_response(output_text, request_id))
-
-            except Exception as e:
-                self.logger.log_error(f"[RID: {request_id}] Error during execution: {e}")
-                err_tensor = pb_utils.Tensor("text_output", np.array([str(e).encode('utf-8')], dtype=np.bytes_))
-                responses.append(pb_utils.InferenceResponse(output_tensors=[err_tensor]))
-
-        return responses
-
-    def _parse_input(self, request) -> Tuple[Union[str, List[Dict]], bool]:
-        input_text = self._get_input_scalar(request, "text_input")
-        try:
-            conversation = json.loads(input_text)
-            if isinstance(conversation, list):
-                return conversation, True
-        except (json.JSONDecodeError, TypeError):
-            pass
-        return input_text, False
-
-    def _tokenize(self, input_data, is_chat: bool):
-        if self.supports_chat_template and is_chat:
-            return self.tokenizer.apply_chat_template(
-                input_data,
-                tokenize=True,
-                add_generation_prompt=True,
-                return_tensors="pt",
-                return_dict=True
-            ).to(self.device)
-        else:
-            if is_chat:
-                 input_data = str(input_data)
-            return self.tokenizer(input_data, return_tensors="pt").to(self.device)
-
-    def _generate(self, inputs, gen_config: GenerationConfig) -> str:
-        input_ids = inputs["input_ids"]
-        input_len = input_ids.shape[-1]
-
-        with torch.no_grad():
-            outputs = self.model.generate(
-                **inputs,
-                generation_config=gen_config,
-                pad_token_id=self.tokenizer.pad_token_id,
-                eos_token_id=self.tokenizer.eos_token_id
-            )
-
-        generated_tokens = outputs[0][input_len:]
-        decoded_output = self.tokenizer.decode(generated_tokens, skip_special_tokens=True)
-        return decoded_output.strip()
-
-    def _create_generation_config(self, request) -> GenerationConfig:
-        def get_param(name, default=None, cast_type=None):
-            val = self._get_input_scalar(request, name, default)
-            if val is not None and cast_type:
-                return cast_type(val)
-            return val
-
-        return GenerationConfig(
-            max_length=get_param("max_length", 1024, int),
-            max_new_tokens=get_param("max_new_tokens", 256, int),
-            temperature=get_param("temperature", 1.0, float),
-            do_sample=get_param("do_sample", False, bool),
-            top_k=get_param("top_k", 50, int),
-            top_p=get_param("top_p", 1.0, float),
-            repetition_penalty=get_param("repetition_penalty", 1.0, float),
-        )
-
-    def _create_response(self, output_text: str, request_id: str):
-        """생성된 텍스트를 Triton Response 객체로 변환"""
-        output_tensor = pb_utils.Tensor(
-            "text_output",
-            np.array([output_text.encode('utf-8')], dtype=np.bytes_)
-        )
-        return pb_utils.InferenceResponse(output_tensors=[output_tensor])
-
-    def _get_config_param(self, key: str, default: str = None) -> str:
-        params = self.model_config.get('parameters', {})
-        if key in params:
-            return params[key].get('string_value', default)
-        return default
-
-    def _get_input_scalar(self, request, name: str, default=None):
-        tensor = pb_utils.get_input_tensor_by_name(request, name)
-        if tensor is None:
-            return default
-        return self._np_decoder(tensor.as_numpy()[0])
-
-    def _np_decoder(self, obj):
-        if isinstance(obj, bytes):
-           return obj.decode('utf-8')
-        if np.issubdtype(obj, np.integer):
-           return int(obj)
-        if np.issubdtype(obj, np.floating):
-           return round(float(obj), 3)
-        if isinstance(obj, np.bool_):
-           return bool(obj)
-
-    def finalize(self):
-        self.logger.log_info(f"Finalizing model {self.model_name}")
-        self.model = None
-        self.tokenizer = None
-        torch.cuda.empty_cache()

README.md

@@ -0,0 +1,35 @@
+---
+license: apache-2.0
+inference: false 
+base_model: Qwen/Qwen3-4B-Instruct
+base_model_relation: quantized 
+tags: [green, llmware-chat, p4, gguf,emerald]
+---
+
+# qwen3-4b-instruct-gguf
+
+**qwen3-4b-instruct-gguf** is a GGUF Q4_K_M int4 quantized version of [Qwen3-4B-Instruct](https://www.huggingface.co/Qwen/Qwen3-4B-Instruct), providing a very fast inference implementation, optimized for AI PCs.    
+
+This is from the latest release series from Qwen, and has 'thinking' capability expressed as 'think' tokens.  
+
+This model will run on an AI PC with at least 16 GB of memory.   
+
+### Model Description
+
+- **Developed by:** Qwen
+- **Model type:** qwen3
+- **Parameters:** 4 billion  
+- **Model Parent:** Qwen/Qwen3-4B-Instruct  
+- **Language(s) (NLP):** English  
+- **License:** Apache 2.0  
+- **Uses:** Chat, general-purpose LLM  
+- **Quantization:** int4  
+  
+
+## Model Card Contact  
+
+[llmware on github](https://www.github.com/llmware-ai/llmware) 
+
+[llmware on hf](https://www.huggingface.co/llmware)  
+
+[llmware website](https://www.llmware.ai)  

config.json

@@ -0,0 +1,30 @@
+{
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2560,
+  "initializer_range": 0.02,
+  "intermediate_size": 9728,
+  "max_position_embeddings": 40960,
+  "max_window_layers": 36,
+  "model_type": "qwen3",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.0",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

config.pbtxt

@@ -1,135 +0,0 @@
-# Triton Backend for TransformerLLM.
-backend: "python"
-max_batch_size: 0
-
-# Triton should expect as input a single string
-# input of variable length named 'text_input'
-input [
-
-  {
-    name: "text_input"
-    data_type: TYPE_STRING
-    dims: [ 1 ]
-
-
-  },
-  {
-    name: "max_length"
-    data_type: TYPE_INT32
-    dims: [ 1 ]
-
-    optional: true
-
-
-  },
-  {
-    name: "max_new_tokens"
-    data_type: TYPE_INT32
-    dims: [ 1 ]
-
-    optional: true
-
-
-  },
-  {
-    name: "do_sample"
-    data_type: TYPE_BOOL
-    dims: [ 1 ]
-
-    optional: true
-
-
-  },
-  {
-    name: "top_k"
-    data_type: TYPE_INT32
-    dims: [ 1 ]
-
-    optional: true
-
-
-  },
-  {
-    name: "top_p"
-    data_type: TYPE_FP32
-    dims: [ 1 ]
-
-    optional: true
-
-
-  },
-  {
-    name: "temperature"
-    data_type: TYPE_FP32
-    dims: [ 1 ]
-
-    optional: true
-
-
-  },
-  {
-    name: "repetition_penalty"
-    data_type: TYPE_FP32
-    dims: [ 1 ]
-
-    optional: true
-
-
-  },
-  {
-    name: "stream"
-    data_type: TYPE_BOOL
-    dims: [ 1 ]
-
-    optional: true
-
-
-  }
-
-]
-
-
-# Triton should expect to respond with a single string
-# output of variable length named 'text_output'
-output [
-
-  {
-    name: "text_output"
-    data_type: TYPE_STRING
-    dims: [ 1 ]
-
-  }
-
-]
-
-parameters: [
-  {
-    key: "base_model_path",
-    value: {string_value: "/cheetah/input/model/groupuser/qwen3-4b-instruct-gguf"}
-  },
-  {
-    key: "gguf_filename",
-    value: {string_value: "Qwen3-4B-Q4_K_M.gguf"}
-  },  
-  {
-    key: "is_adapter_model",
-    value: {string_value: "false"}
-  },
-  {
-    key: "adapter_model_path",
-    value: {string_value: ""}
-  },
-  
-  {
-    key: "quantization",
-    value: {string_value: "none"}
-  }
-]
-
-instance_group [
-  {
-    kind: KIND_AUTO
-    count: 1
-  }
-]
-

hash_record_sha256.json

@@ -0,0 +1,4 @@
+{
+ "Qwen3-4B-Q4_K_M.gguf": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
+ "time_stamp": "2025-07-05_070138"
+}